beyoru/Kiwen-27B
Kiwen-27B is a 27 billion parameter language model developed by beyoru, fine-tuned from Qwen3.8-27B with a 32K context length. It is specifically optimized to improve reasoning task completion by ensuring the model terminates its reasoning and produces a final answer within its generation budget. This fine-tuning significantly enhances performance on benchmarks like GSM8K by teaching the model when to stop deliberating and provide a concrete output.
Loading preview...
Kiwen-27B: Optimized for Reasoning Completion
Kiwen-27B is a 27 billion parameter model, fine-tuned from the robust Qwen3.8-27B base. Its primary innovation lies in its training on long chain-of-thought traces from Kimi K3, specifically to teach the model to terminate its reasoning and produce an answer before exhausting its generation budget. This addresses a common issue where strong reasoning models might deliberate extensively without delivering a final output.
Key Capabilities & Performance
- Improved Reasoning Completion: The model demonstrates a significant improvement in producing final answers for complex reasoning tasks. On the GSM8K benchmark, it shows a +10.31 percentage-point improvement under flexible extraction, primarily by learning when to stop and output an answer.
- Consistent Gains: While the largest gains are on GSM8K, other benchmarks like IFEval and VMLU also show smaller but consistently positive improvements, with no measured regressions in core capabilities.
- Efficiency in Output: The fine-tuning mechanism focuses on behavioral change, ensuring the model efficiently delivers an answer rather than solely enhancing its fundamental reasoning ability.
When to Use Kiwen-27B
- Reasoning-heavy tasks: Ideal for applications where a definitive, timely answer is crucial, especially in mathematical or logical problem-solving contexts.
- Budget-constrained generation: Useful when working with limited token generation budgets, as it's designed to provide an answer within those constraints.
- Building on Qwen3.8-27B: If you are already considering Qwen3.8-27B for its reasoning strengths, Kiwen-27B offers an enhanced version focused on practical output delivery.
Limitations
- Not broadly stronger: This fine-tune primarily targets output completion behavior, not a fundamental increase in reasoning power across all domains. Out-of-domain evaluations showed a slight, though not statistically significant, increase in negative log-likelihood on unrelated tasks, suggesting minimal impact on general capabilities.